NVIDIA's open transcription model labels speakers and takes prompts
Sorted AI30 OctNVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31 of them to or from English and labels up to eight speakers (speaker labels work in 13 languages). It handles recorded audio up to an hour long, not live streams, and you can steer it with plain instructions such as a list of terms to spell correctly, or ask it questions about a meeting. NVIDIA says it ranks first among open-weight models on the Hugging Face Open ASR leaderboard for English, and in a case study on recordings of Indian doctors, fine-tuning cut word errors on unseen speakers from 22.44% to 12.17%. The leaderboard and case study results are NVIDIA's own reporting.
Why it matters: If you transcribe sales calls, client meetings or dictation, you can try it in the Hugging Face demo or through hosted endpoints like Baseten and DeepInfra, and test a prompt with your own product or client names. Spot-check names and speaker changes against the recording before relying on it.
NVIDIA (Hugging Face blog) ↗AI-assisted, reviewed by Sorted AI
▲ 0
0 comments